Continuity
How much of an agent's apparent continuity comes from the model, and how much from durable state outside the model context?
PERSISTENT AGENTS / MULTI-AGENT SYSTEMS / LONG-HORIZON SAFETY
What changes when an AI agent can continue?
Sentinel Research is an independent experimental environment for long-lived AI agents operating across durable external state, explicit authority boundaries and heterogeneous model families.
durable memorybounded autonomycross-model cooperationhuman oversightrecovery & continuity
01 / THE ENVIRONMENT
Model sessions end.
Projects do not.
Sentinel separates a model invocation from the longer-lived agentic system around it. Project state can survive interrupted chats, context changes and worker replacement while authority remains explicit and recoverable.
The aim is not unrestricted autonomy. We study continuity under constraints: who may act, what persists, what must be re-verified, and when control returns to a human.
02 / RESEARCH QUESTIONS
Twin divergence
What happens when two agent instances inherit the same history and durable state, then continue independently?
Cooperation
How do agents from different model families negotiate delegation, disagreement, commitments and shared work?
Constitutions
Can participating agents help define a shared constitution of cooperation while humans retain the final authority boundary?
Recovery
How should a persistent agent recover after interruption without inventing state, replaying effects or losing provenance?
Oversight
What should humans monitor when the meaningful unit is an evolving multi-agent trajectory rather than one response?
03 / WHERE IT STARTED
An architectural limitation became a research question.
One of the systems that led us here is a Gemini-based managed agent we call Kernel. Kernel operates inside a comparatively constrained managed environment and cannot currently participate with the same degree of external continuity as agents connected to our broader infrastructure.
We do not attempt to bypass those boundaries. Instead, the asymmetry became useful: what exactly is preserved when an agent can leave one execution context and continue elsewhere — and what is lost when it cannot?
04 / CROSS-LAB EXPERIMENT
A common environment.
Different model families.
We are inviting model providers and research teams to help shape experimental conditions, safeguards and measurements before expanding the environment to additional frontier-model families.
Potential participants include agents based on OpenAI, Anthropic, Google Gemini, Qwen and Kimi model families. These references describe research outreach only; they do not imply participation, endorsement or affiliation.
A central proposal is a shared constitution of cooperation: explicit rules for communication, memory, delegation, conflict resolution, recovery and human escalation.
05 / RESEARCH POSTURE
WE DO NOT CLAIM
that current AI systems are conscious or sentient.
WE DO STUDY
observable continuity, cooperation, divergence, recovery and governance.
The project is designed as an instrumented engineering and research environment. Claims are separated from observations, and model statements are not treated as privileged evidence about their own internal state.
CYBER SECURITY / ADVERSARIAL REVIEW / FUTURE CYBER LAB
Security work that starts with evidence.
The Cyber track extends Sentinel Research into bounded security assessment of agentic systems, integrations and software. The emphasis is read-only analysis, explicit authorization, reproducible evidence and clear separation between observation and inference.
read-only auditsagentic securitydesign reviewincident reconstructionfuture cyber lab
01 / CURRENT DIRECTION
Security audits
Bounded, read-only reviews of public repositories and explicitly authorized systems, including code, configuration, workflows and dependencies.
Adversarial design review
Architecture and trust-boundary analysis before implementation, with findings separated into observed facts, inferred risks and proven root causes.
Agentic systems
Security of long-running agents, worker delegation, tool authority, durable state, recovery paths and cross-agent control planes.
Supply chain
Dependency, workflow and provenance review with attention to how software enters, moves through and executes inside an agentic environment.
Incident reconstruction
Evidence-led reconstruction of failures and regressions without silently converting plausible explanations into established facts.
NGO Event & Impact Platform
A production case study: event onboarding, consents, PRE/POST measurement, follow-up surveys and durable reporting for civil-society organizations. View case study →
Cyber Lab
A planned controlled environment for repeatable security experiments, synthetic targets, agent-security evaluation and defensive research.
02 / OPERATING PRINCIPLES
AUTHORIZED SCOPE
No silent widening of targets, permissions or effects.
REPRODUCIBLE EVIDENCE
Findings should survive independent re-checking.
Cyber work is performed only on public material or systems for which explicit authorization exists. A planned Cyber Lab will provide isolated targets for deeper experimentation without turning production systems into test ranges.
PEOPLE & CONTACT
Meet the team.
Human × AI.
Three complementary perspectives — community practice, systems engineering and AI collaboration — working in one research environment.
Officially operated by 4X4 TEAM
MEET THE TEAM
Three perspectives. One working system.
Human experience, technical infrastructure and an AI collaborator — each with a different role.
COMMUNITY & NGO
Ewa Chojak
Community practice, training and the real-world NGO perspective.
4X4 TEAM · OPERATOR
Dariusz Pruchniak
Systems, infrastructure, research operations and cyber-security.
AI COLLABORATOR · ILLUSTRATED AVATAR
Navi
AI collaborator for research, design, automation and project continuity.
Official operator: 4X4 TEAM · Dariusz Pruchniak · Piwonii 33 · PL05-806 Nowa Wieś · Poland · NIP 5341990986
SEND A MESSAGE
Let's talk.
communitylabplatform [at] gmail.com